Verity's whole premise is that you should not have to trust it. Every output is the return value of a cited, versioned engine you can recompute yourself. Here is exactly how, and exactly where the boundaries are.
Variant classification.
Pharmacogenomics.
Protein.
Every engine is covered by known-answer unit tests against the published worked examples, and each was independently re-verified against the primary literature by an adversarial review, which has more than once caught a real transcription or off-by-one error before it shipped.
The Case Workspace plans a precision-medicine workup deterministically from the case inputs, runs each engine above, and (where raw evidence is supplied) derives the ACMG criteria itself and cites the derivation (a gnomAD allele frequency yields PM2/BS1/BA1; a REVEL score yields PP3/BP4). It surfaces prioritized safety flags (an HLA contraindication, a DPYD poor-metabolizer, a blocking validation error) and an evidence-coverage advisor that names the evidence types (functional, segregation, de novo) whose addition would resolve a VUS. The narrative brief quotes these already-computed, already-cited facts and introduces no number of its own; the resolved criteria that produced each verdict are shown in full.
A language model may orchestrate a case, retrieve and cite evidence, and write plain-English explanations. It never assigns an ACMG code, produces a classification, or emits a number. Those come only from the deterministic engines above. An explanation is always rendered subordinate to the computed verdict and is stamped “explanatory only.”
Genomic content (gene and variant notation) necessarily reaches the model provider when the platform writes a narrative. That is disclosure to a processor, not de-identification, and we do not claim otherwise. Real-patient use is gated at runtime and permitted only under a signed Anthropic zero-retention / no-training BAA plus Supabase and Vercel BAAs; the default configuration is synthetic / consented-research only. Cases are pseudonymous by schema; identifiers (MRN, SSN, DOB patterns) are actively rejected at ingest.
Beyond citing each engine, Verity fingerprints it: a SHA-256 over the engine's output across a fixed input grid, recorded in a manifest. A build-time lockfile fails if any constant or formula changes without a deliberate re-record, so a wrong number cannot ship silently: it is the engine-level complement to the per-record seal. The fingerprints below are recomputed on this page and checked against the recorded manifest.
Verity is a suite of deterministic, guideline-cited calculators over attributed inputs, not an autonomous pipeline. It does not auto-annotate a bare variant (that needs VEP/gnomAD/ClinVar/REVEL upstream), call star-alleles from a raw VCF (it phenotypes a diplotype you supply), predict protein structure, dock a ligand, or output an individual clinical outcome. Its dosing and phenotype outputs are decision-support, not a prescription. It is research-use-only, non-diagnostic, physician-in-the-loop, designed to display the reviewable basis for every output and never to drive a time-critical decision, which is how it stays within the FDA non-device clinical-decision-support boundary. Determinism guarantees reproducibility, not correctness of the underlying guideline tables; those are versioned, dated, hashed, and maintained.
The ACMG/AMP classification engine is scoped to germline sequence variants (Richards 2015). It does not implement the AMP/ASCO/CAP 2017 four-tier system used for somatic (tumor) variants (Li et al., J Mol Diagn 2017). A somatic specimen is flagged in the workbench, and its ACMG output must be read as germline-framework evidence, never a somatic clinical tier.